CN108875532A - A kind of video actions detection method based on sparse coding and length posterior probability - Google Patents
A kind of video actions detection method based on sparse coding and length posterior probability Download PDFInfo
- Publication number
- CN108875532A CN108875532A CN201810073174.9A CN201810073174A CN108875532A CN 108875532 A CN108875532 A CN 108875532A CN 201810073174 A CN201810073174 A CN 201810073174A CN 108875532 A CN108875532 A CN 108875532A
- Authority
- CN
- China
- Prior art keywords
- segment
- movement
- classification
- length
- training
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/20—Movements or behaviour, e.g. gesture recognition
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
- G06F18/2135—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/28—Determining representative reference patterns, e.g. by averaging or distorting; Generating dictionaries
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
- G06V20/46—Extracting features or characteristics from the video content, e.g. video fingerprints, representative shots or key frames
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
- G06V20/49—Segmenting video sequences, i.e. computational techniques such as parsing or cutting the sequence, low-level clustering or determining units such as shots or scenes
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Multimedia (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computing Systems (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Psychiatry (AREA)
- Social Psychology (AREA)
- Human Computer Interaction (AREA)
- Image Analysis (AREA)
Abstract
The video actions detection method based on sparse coding and length posterior probability that the present invention provides a kind of, including feature extraction, generation movement segment are proposed, movement segment proposes that classification, redundancy detection remove Four processes；Firstly, extracting the output of the last one full articulamentum of network as feature by video input into Three dimensional convolution neural network C3D network；Then proposed using context-sensitive sparse coding method generation movement segment；Subsequently classified using classifier to movement segment proposal, classification results are adjusted using length posterior probability after classification；Finally redundancy detection is removed using non-maxima suppression method.The present invention is proposed using context-sensitive sparse coding generation movement segment, the movement segment that can obtain the high quality comprising most realistic operation segments is proposed, and classification results are adjusted using length posterior probability after movement segment proposes classification, the precision of motion detection result can be greatly improved.
Description
Technical field
The present invention relates to video human motion detection technologies in a kind of computer vision field, especially a kind of based on sparse
The video actions detection method of coding and length posterior probability.
Background technique
In video human action detection be the non-editing at one section long video in detect one movement since when
To when terminating and identify the classification of the movement.Motion detection is widely used in intelligent video monitoring, smart home, man-machine friendship
Mutually, motion analysis, virtual reality etc..However, in face of the massive video generated daily, this task allows the mankind to carry out
Words will be very inefficient and uninteresting, therefore, using information useful in Computer Automatic Extraction video be very there is an urgent need to.
Human action detection in video is divided into two key steps of detection that human action is indicated and acted, wherein moving
The detection of work is divided into the proposal of generation movement segment again, movement segment proposes that classification and redundancy detection remove three steps.With
The continuous development of computer vision technique, the research in terms of motion detection are increasingly taken seriously, various motion detection method layers
It is not poor out.In terms of conventional method, achieved using the methods of sparse coding, random forest, segment bag progress motion detection outstanding
Achievement；In recent years, with the continuous development of deep learning, the methods of convolutional neural networks and shot and long term memory network are used
Carrying out motion detection also made breakthrough progress.
Up to the present, although having had already appeared many outstanding methods and acquisition in the research of human action detection algorithm
It is outstanding as a result, still needing to solve there are still some problems.First, the method that existing generation movement segment is proposed exists
When selecting candidate segment, while removing non-action segment, many movement segments can be also eliminated, are caused many dynamic
It loses.The movement segment of high quality proposes to need to include movement segment as much as possible under the premise of quantity is few as far as possible, this
When lose excessive movement and will affect the precision of last motion detection result.Second, when classifying to movement segment proposal,
It has some lesser segments Chong Die with realistic operation segment and possesses very high classification score, for example some movement segments are proposed only
It is the sub-fraction in correct movement segment, these segments are not correct testing result, but since it does not include background
So classification score is very high, which results in subsequent non-maxima suppression algorithm, the segment of these mistakes can inhibit correct
The larger lower segment of unit fraction Chong Die with realistic operation segment, eventually lead to motion detection result mistake.
Summary of the invention
The video actions detection method based on sparse coding and length posterior probability that the purpose of the present invention is to provide a kind of,
Propose including feature extraction, generation movement segment, movement segment proposes that classification and redundancy detection remove Four processes：
Characteristic extraction procedure includes the following steps：
Step 101, in training set video and test video input C3D network, the input of C3D network first tier is view
16 frame images in frequency extract the last one full articulamentum of network using every 16 frame of video as in a slice input network
Output is used as feature；
Step 102, feature will be obtained in step 101 carries out dimensionality reduction using principal component analysis；
Step 103, the feature after dimensionality reduction in step 102 is normalized feature using L2 norm；
Generation movement segment proposes that process includes the following steps：
Step 201, training set video be cut into realistic operation segment and with the friendship of realistic operation segment and exist than IoU
A certain range of segment；
Step 202, using the video clip feature sheared in step 201, respectively each realistic operation segment and every
A friendship with realistic operation segment and the piecemeal learning sparse dictionary than IoU in a certain range；
Step 203, test video is sheared using sliding time window method, generates candidate segment；
Step 204, candidate segment is reconstructed respectively using the dictionary learnt in step 202, and calculates reconstructed error；
Step 205, the reconstructed error in conjunction with obtained in step 204 obtains each word using the non-maxima suppression method of weighting
The movement segment that allusion quotation is calculated is proposed；
Step 206, the movement segment proposal of dictionary creation each in step 205 is combined, then is once weighted
Non-maxima suppression inhibit method obtain final movement segment proposed issue；
Movement segment proposes that assorting process includes the following steps：
Step 301, using one classification classifier of movement v.s. non-action two of training set video training, and training one
Multi-class classifier；
Step 302, the final action movie that will be generated in step 206 using two classification classifiers of training in step 301
Duan Tiyi carries out two classification, and removal is classified as the proposal of non-action；
Step 303, the proposal remained in step 302 is carried out using more classification classifiers of training in step 301
Classification；
Step 304, length posterior probability of all categories is calculated using training set；
Step 305, classification results obtained in step 303 are carried out using length posterior probability obtained in step 304
Adjustment；
Redundancy detection removal process includes the following steps：
Step 401, propose that new score carries out non-maxima suppression algorithm meter using movement segment obtained in step 305
It calculates, removal redundancy detection obtains motion detection result.
Compared with prior art, the present invention having the following advantages that：(1) it is used with traditional sparse coding when learning dictionary
The movement segment sheared is different, and the present invention uses pure movement segment incessantly, is additionally added some comprising movement fractional time
The relevant information of context, i.e., the segment comprising certain background before and after movement segment are reached with reinforcing the generalization ability of dictionary
Raising acts the effect that segment proposes quality；(2) present invention has first used a movement v.s. non-action two before more classification
Classifier further screens out non-action segment, can reduce the calculation amount of subsequent operation and improve motion detection precision；
(3) third, the present invention is after classifying to movement segment proposal, classification score is adjusted using length posterior probability
It is whole, reach the score of reduction lesser segment be overlapped with realistic operation segment, improves biggish be overlapped with realistic operation segment
The purpose of the score of section, to improve the precision of motion detection.
The invention will be further described with reference to the accompanying drawings of the specification.
Detailed description of the invention
Fig. 1 is the method for the present invention flow diagram.
Fig. 2 is C3D network structure.
Fig. 3 is that movement segment proposes example schematic diagram.
Fig. 4 is that movement segment proposes two classification process schematic diagrames.
Specific embodiment
The present invention proposes a kind of motion detection method based on sparse coding and length posterior probability, including feature extraction,
Generation acts that segment is proposed, movement segment proposes that classification and redundancy detection remove Four processes, to the long video of non-editing into
The a series of calculating of row obtains at the beginning of wherein human action occurs, the classification of end time and movement.Video actions
The basic framework of detection technique is as shown in Figure 1, the present invention is carried out according to this basic framework.
Characteristic extraction procedure includes the following steps：
For step 1) in training video and test video input C3D network, C3D network structure is as shown in Figure 2.C3D net
The input of network first layer is 16 frame images in video, is inputted in network using every 16 frame of video as a slice, the (1~
15), (2~16) ... frame as input, then extract the last one full articulamentum fc7 layers output of network as feature, it is defeated
It is out 4096 dimensions.Then, if the frame number of video is F, the feature of video is (F-15) × 4096 dimension.
Step 2) will obtain feature and carry out dimensionality reduction using principal component analysis in step 1), drop to 500 dimensions from 4096 dimensions.
Feature after dimensionality reduction in step 2) is normalized step 3) using L2 norm.
Generation movement segment proposes that process includes the following steps：
Step 4) the video in training set be cut into realistic operation segment and with the friendship of realistic operation segment and ratio
(IoU) segment in (0.6~0.7), (0.7~0.8), (0.8~0.9), (0.9~1) this four sections.
Step 5) using the video clip feature sheared in step 4), be realistic operation segment and with realistic operation segment
Friendship and than the segment of (IoU) in (0.6~0.7), (0.7~0.8), (0.8~0.9), (0.9~1) this four sections point
Not Xue Xi 5 sparse dictionaries, i.e., context-sensitive sparse dictionary.The specific method is as follows for dictionary learning：
XiIndicate the feature for being used to the video clip of training dictionary, X=[X1|…|Xi], i is for trained all segments
The characteristic for being included.The study of dictionary D is carried out by solving following formula：
Wherein, A is rarefaction representation coefficient matrix；D is the dictionary to be learnt；Y is the classification mark of feature used in training
Label, each C3D feature vector are owned by a class label；W is an one-to-many classifier；Coefficient lambda1、λ2、λ3Respectively
It is 0.05,0.05,0.025；N is the quantity of segment characterizations used in training.The study of dictionary is the process of an iteration,
It each time in iteration, using the strategy alternately updated, first fixes A and updates D, W, update A in fixed D, W, finally obtaining makes formula
(1) the smallest result.It updates W and uses following formula：
It updates A and uses following formula：
It updates D and uses following formula：
After reaching iteration ending standard, sparse dictionary D required for us is obtained.Iteration ending standard be twice iteration it
Between difference less than 0.01, or more than maximum number of iterations 300 times.
Step 6) shears test video using sliding time window method, generates candidate segment.Sliding time window
In length of window the realistic operation fragment length in training set cluster using average drifting (Meanshift) algorithm
It arrives.
Step 7) is using the context-sensitive sparse dictionary learnt in step 5) respectively to candidate obtained in step 6)
Segment is encoded, and reconstructed error is calculated, and the score of candidate segment is calculated according to reconstructed error.Circular is as follows：
XkIndicate a candidate segment, coding passes through following formula：
Wherein nkFor the quantity of candidate segment feature；Coefficient lambda is 0.05.After coding, the reconstruct for calculating candidate segment is missed
The calculating of difference, reconstructed error cost is as follows：
Reconstructed error is smaller, illustrates in this segment that a possibility that including movement is bigger, in order to eliminate otherness, counterweight
Structure error is normalized, and obtains the score of candidate segment, and the calculating of score score is as follows：
Step 8) combines candidate segment score obtained in step 7) to inhibit using the non-maxima suppression of weighting
Movement segment, which is calculated, in (WeightedNon-Maximum Suppression, WNMS) algorithm proposes, movement segment proposal is shown
Such as Fig. 3.Different from common NMS method, WNMS uses different inhibition overlap coefficients for the segment of different length.For
Fragment length is (0~50), (50~500), (500~600), (600~800), (800~1200), (1200~video is long
Degree) range segment, use respectively 0.55,0.65,0.55,0.45,0.35,0.25 as inhibition overlap coefficient.
Step 9) combines the movement segment proposal of dictionary creation each in step 8), then carries out a WNMS and obtain
Final movement segment proposed issue.So-called combination is exactly all added together and then is removed the result of each dictionary is simple
Wherein duplicate part.
Movement segment proposes that assorting process includes the following steps：
Step 10) uses one classification classifier of movement v.s. non-action two of training set video training and a multiclass
Other classifier.Specific training step is as follows：
Step 10-1) the classification SVM classifier of training action v.s. non-action two.Using true movement segment as positive training
Collect Xaction, the pure background segment that the IoU with true movement segment is 0 is as negative training set Xback, and pass through random selection
Background segment is come essentially identical, the i.e. N that guarantees the quantity of negative training set and the quantity of positive training setaction≈Nback.SVM points of training
The parameter error item penalty C that class device uses is set as 1.0, and kernel function uses radial basis function (Radial
BasedFunction, RBF), the parameter gamma in RBF kernel function is set as 0.2.Video in training set is all a whole segment length
Video, wherein having movement and background, training is concentrated with the position of mark movement in video, and the video clip not acted is exactly
Background.It is exactly the segment for being 0 with movement segment IoU.
Step 10-2) one one-to-many SVM classifier of training.It is different from two classification for balance training collection quantity, more points
Class will reduce the quantity of background, be allowed to essentially identical with the quantity of each movement class.SVM parameter setting and step 10-1) in
It is identical.
The movement segment proposal that step 11) will be generated using two classification classifiers of training in step 10-1) in step 9)
Two classification are carried out, each movement segment proposes to be sliced comprising multiple 16 frame video features, each is sliced after sorting
Be owned by a classification results, if after sorting, a movement segment propose in have more than 30% slice and be classified as carry on the back
Scape, then it is considered that this movement segment proposes to be background, otherwise it is considered that this movement segment is proposed as movement, such as Fig. 4
It is shown.The proposal for being only classified as movement, which can remain, continues subsequent calculating.When calculating C3D feature, C3D network the
One layer of input is 16 frame images in video, using every 16 frame of video as in a slice input network, then extracts network
The last one full articulamentum fc7 layers output is exported as feature as 4096 dimensions.One slice is exactly the feature of 16 frames.
Step 12) carries out the proposal remained in step 11) using more classification classifiers of training in step 10-2)
Classification.The classification that movement segment is proposed is that wherein most that occurs in all be sliced in the classification that is classified into, action movie
The score that the probability value of Duan Tiyi, i.e. this movement segment are proposed, the classification of this candidate segment is classified into for all slices
Probability average value.
Step 13) calculates length posterior probability of all categories using training set.Circular is as follows：
One fixed length S is set first, some length T={ S, 2S, 3S ... } are set by step-length of S, then dynamic
Make length L to be categorized into these regular lengths T, then calculate the posterior probability of each length T, calculates and use following formula：
Wherein, ∑ LTFor the quantity of all movement length L being categorized into this length T, ∑ L is everything length L
Quantity.Probability P is length posterior probability, i.e., the probability that the movement segment of this length occurs in this classification.
Step 14) adjusts classification results obtained in step 12) using length posterior probability obtained in step 13)
Whole, circular is as follows：
The score that movement segment is proposed is adjusted using naive Bayesian posterior probability, formula used is as follows：
Wherein, P (L | Ci) it is movement length posterior probability in the i-th class being calculated in step 13).P(Ci|f,Θ)
It is the classification score proposed in step 12) using the movement segment that SVM classifier obtains, this classification score is exactly this movement
Segment proposes the probability for being categorized into the i-th class, and wherein f is C3D feature used in SVM, and Θ is the parameter in SVM.P (L) is
Prior probability, it is 1 that it, which is arranged,.P(Ci| f, Θ, L) it is the new score that the movement segment obtained by adjusting after is proposed.
Redundancy detection removal process includes the following steps：
Step 15) proposes that new score carries out non-maxima suppression (Non- using movement segment obtained in step 14)
Maximum Suppression, NMS) algorithm calculating, it removes redundancy detection and obtains final motion detection result.It is arranged in NMS
Duplication threshold alpha be slightly smaller than mean accuracy mean value when testing result is evaluated (mean Average Precision, mAP)
The Duplication threshold θ (α=θ -0.1) used.
Claims (10)
1. a kind of video actions detection method based on sparse coding and length posterior probability, including feature extraction, generation movement
Segment is proposed, movement segment proposes classification and redundancy detection removes Four processes：
Characteristic extraction procedure includes the following steps：
Step 101, in training set video and test video input C3D network, the input of C3D network first tier is in video
16 frame images extract the output of the last one full articulamentum of network using every 16 frame of video as in slice input network
As feature；
Step 102, feature will be obtained in step 101 carries out dimensionality reduction using principal component analysis；
Step 103, the feature after dimensionality reduction in step 102 is normalized feature using L2 norm；
Generation movement segment proposes that process includes the following steps：
Step 201, training set video is cut into realistic operation segment and with the friendship of realistic operation segment and than IoU certain
Segment in range；
Step 202, using the video clip feature sheared in step 201, respectively each realistic operation segment and each with
The friendship of realistic operation segment and piecemeal learning sparse dictionary than IoU in a certain range；
Step 203, test video is sheared using sliding time window method, generates candidate segment；
Step 204, candidate segment is reconstructed respectively using the dictionary learnt in step 202, and calculates reconstructed error；
Step 205, the reconstructed error in conjunction with obtained in step 204 obtains each dictionary meter using the non-maxima suppression method of weighting
Obtained movement segment is proposed；
Step 206, the movement segment proposal of dictionary creation each in step 205 is combined, then is once weighted non-
Maximum inhibits inhibition method to obtain final movement segment proposed issue；
Movement segment proposes that assorting process includes the following steps：
Step 301, using one classification classifier of movement v.s. non-action two of training set video training, and one multiclass of training
Other classifier；
Step 302, the final movement segment generated in step 206 is mentioned using two classification classifiers of training in step 301
View carries out two classification, and removal is classified as the proposal of non-action；
Step 303, classified using more classification classifiers of training in step 301 to the proposal remained in step 302；
Step 304, length posterior probability of all categories is calculated using training set；
Step 305, classification results obtained in step 303 are adjusted using length posterior probability obtained in step 304；
Redundancy detection removal process includes the following steps：
Step 401, propose that new score carries out the calculating of non-maxima suppression algorithm using movement segment obtained in step 305,
Removal redundancy detection obtains motion detection as a result, mean accuracy is equal when wherein Duplication threshold alpha is evaluated less than testing result
The Duplication threshold θ that value uses, α=θ -0.1.
2. the method according to claim 1, wherein the video in training set is cut into really in step 201
Act segment and with the friendship of realistic operation segment and than IoU in (0.6~0.7), (0.7~0.8), (0.8~0.9), (0.9
~1) segment in this four sections.
3. according to the method described in claim 2, it is characterized in that, the step 202 detailed process is as follows：
Step 2021, using XiIndicate the feature of the video clip of training dictionary, X=[X1|…|Xi], i is for trained institute
There is the characteristic that segment is included, carries out the study of dictionary D by solving formula (1)：
Wherein, A is rarefaction representation coefficient matrix, and D is the dictionary to be learnt, and W is an one-to-many classifier, and Y is training institute
The class label of the feature used, coefficient lambda1、λ2、λ3Respectively 0.05,0.05,0.025, n be training used in segment characterizations
Quantity, F refers to F norm.
Step 2022, formula (1) is iterated using the strategy alternately updated, first fixes A and updates D, W, update A in fixed D, W,
Wherein
It updates W and uses following formula：
It updates A and uses following formula：
It updates D and uses following formula：
It step 2023, if the difference between iteration is less than 0.01 twice, or is more than maximum number of iterations, iteration stopping is chosen
Corresponding dictionary sparse dictionary when making the result minimum of formula (1).
4. the method according to claim 1, wherein the length of window in step 203 in sliding time window makes
The realistic operation fragment length in training set is clustered to obtain with mean shift algorithm.
5. the method according to claim 1, wherein the reconstructed error cost in the step 204 is
Wherein, XkIndicate a candidate segment；nkFor the quantity of candidate segment feature；
It is encoded by formula (5)：
Wherein, coefficient lambda 0.05.
6. according to the method described in claim 5, it is characterized in that, the detailed process of step 205 is：
Step 2051, different inhibition overlap coefficients is used for the segment of different length；
Step 2052, retain in different fragments and be greater than maximum corresponding of value in the corresponding score score for inhibiting overlap coefficient
Section, the segment are to act segment to propose
Wherein, min (cost), max (cost) are respectively maximum value, the minimum value in cost.
7. the method according to claim 1, wherein the classification of training action v.s. non-action two SVM in step 301
In classifier：
Using true movement segment as positive training set Xaction, the pure background segment for being 0 the IoU with true movement segment
As negative training set Xback, and the quantity of negative training set and the quantity of positive training set are guaranteed by random selection background segment
It is essentially identical；
The parameter error item penalty C that training SVM classifier uses is set as 1.0, and kernel function uses radial basis function, radial
Parameter gamma in basic function is set as 0.2.
In step 301 in the multi-class classifier of training：
Different from two classification for balance training collection quantity, more classification will reduce the quantity of background, be allowed to act class with each
Quantity it is essentially identical.
8. the method according to the description of claim 7 is characterized in that using two classification point of training in step 301 in step 302
The movement segment generated in step 206 is proposed to carry out two classification, be had more than in each movement segment proposal after classification by class device
30% slice is classified as background, then this movement segment proposes to be background, and otherwise this movement segment is proposed as movement；
The proposal remained in step 302 is divided using the multi-class classifier of training in step 301 in step 303
Class, wherein the classification that movement segment is proposed is that wherein most that occurs in all be sliced in the classification that is classified into, movement
The probability value that segment is proposed is all average values for being sliced and being classified into the probability of classification of this candidate segment.
9. according to the method described in claim 8, it is characterized in that, after step 304 calculates length of all categories using training set
The circular for testing probability is：
One fixed length S is set, length T={ S, 2S, 3S ... } is set by step-length of S, movement length L is categorized into this
In a little regular length T, the posterior probability P of each length T is calculated according to formula (8)：
Wherein, ∑ LTFor the quantity of all movement length L being categorized into this length T, ∑ L is the number of everything length L
Amount.
10. according to the method described in claim 9, it is characterized in that, step 305 is using formula (9) to obtained in step 303 points
Class result is adjusted：
Wherein, P (L | Ci) it is movement length posterior probability in the i-th class being calculated；P(Ci| f, Θ) it is to make in step 303
The classification score proposed with the movement segments that classifiers obtain of classifying, this classification score are exactly that this movement segment proposes to divide more
Class is to the probability of the i-th class, and f is C3D feature used in SVM, and Θ is the parameter in SVM；P (L) is prior probability, it is arranged
It is 1；P(Ci| f, Θ, L) it is the new score that the movement segment obtained by adjusting after is proposed.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810073174.9A CN108875532A (en) | 2018-01-25 | 2018-01-25 | A kind of video actions detection method based on sparse coding and length posterior probability |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810073174.9A CN108875532A (en) | 2018-01-25 | 2018-01-25 | A kind of video actions detection method based on sparse coding and length posterior probability |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108875532A true CN108875532A (en) | 2018-11-23 |
Family
ID=64325977
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810073174.9A Pending CN108875532A (en) | 2018-01-25 | 2018-01-25 | A kind of video actions detection method based on sparse coding and length posterior probability |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108875532A (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109829490A (en) * | 2019-01-22 | 2019-05-31 | 上海鹰瞳医疗科技有限公司 | Modification vector searching method, objective classification method and equipment |
| CN111414827A (en) * | 2020-03-13 | 2020-07-14 | 四川长虹电器股份有限公司 | Depth image human body detection method and system based on sparse coding features |
| CN111918146A (en) * | 2020-07-28 | 2020-11-10 | 广州筷子信息科技有限公司 | Video synthesis method and system |
| CN114216349A (en) * | 2021-12-06 | 2022-03-22 | 浙江大学 | Sintering end point forecasting method based on coding and decoding network |
| CN117352151A (en) * | 2023-12-05 | 2024-01-05 | 吉林大学 | Intelligent accompanying management system and method thereof |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20150325007A1 (en) * | 2012-09-13 | 2015-11-12 | Los Alamos National Security, Llc | Object detection approach using generative sparse, hierarchical networks with top-down and lateral connections for combining texture/color detection and shape/contour detection |
| CN106503659A (en) * | 2016-10-24 | 2017-03-15 | 天津大学 | Action identification method based on sparse coding tensor resolution |
| CN107609509A (en) * | 2017-09-09 | 2018-01-19 | 北京工业大学 | A kind of action identification method based on motion salient region detection |
-
2018
- 2018-01-25 CN CN201810073174.9A patent/CN108875532A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20150325007A1 (en) * | 2012-09-13 | 2015-11-12 | Los Alamos National Security, Llc | Object detection approach using generative sparse, hierarchical networks with top-down and lateral connections for combining texture/color detection and shape/contour detection |
| CN106503659A (en) * | 2016-10-24 | 2017-03-15 | 天津大学 | Action identification method based on sparse coding tensor resolution |
| CN107609509A (en) * | 2017-09-09 | 2018-01-19 | 北京工业大学 | A kind of action identification method based on motion salient region detection |
Non-Patent Citations (2)
| Title |
|---|
| FABIAN CABA HEILBRON 等: "Fast Temporal Activity Proposals for Efficient Detection of Human Actions in Untrimmed Videos", 《2016 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
| XINRAN LIU 等: "Effective Action Detection Using Temporal Context and Posterior Probability of Length", 《INTERNATIONAL CONFERENCE ON MULTIMEDIA MODELING》 * |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109829490A (en) * | 2019-01-22 | 2019-05-31 | 上海鹰瞳医疗科技有限公司 | Modification vector searching method, objective classification method and equipment |
| CN109829490B (en) * | 2019-01-22 | 2022-03-22 | 上海鹰瞳医疗科技有限公司 | Correction vector searching method, target classification method and device |
| CN111414827A (en) * | 2020-03-13 | 2020-07-14 | 四川长虹电器股份有限公司 | Depth image human body detection method and system based on sparse coding features |
| CN111414827B (en) * | 2020-03-13 | 2022-02-08 | 四川长虹电器股份有限公司 | Depth image human body detection method and system based on sparse coding features |
| CN111918146A (en) * | 2020-07-28 | 2020-11-10 | 广州筷子信息科技有限公司 | Video synthesis method and system |
| CN114216349A (en) * | 2021-12-06 | 2022-03-22 | 浙江大学 | Sintering end point forecasting method based on coding and decoding network |
| CN114216349B (en) * | 2021-12-06 | 2022-09-16 | 浙江大学 | Sintering end point forecasting method based on coding and decoding network |
| CN117352151A (en) * | 2023-12-05 | 2024-01-05 | 吉林大学 | Intelligent accompanying management system and method thereof |
| CN117352151B (en) * | 2023-12-05 | 2024-03-01 | 吉林大学 | Intelligent accompanying management system and method thereof |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108875532A (en) | A kind of video actions detection method based on sparse coding and length posterior probability | |
| CN111062382A (en) | Channel pruning method for target detection network | |
| CN111062278B (en) | Abnormal behavior identification method based on improved residual error network | |
| CN110929029A (en) | Text classification method and system based on graph convolution neural network | |
| WO2019179403A1 (en) | Fraud transaction detection method based on sequence width depth learning | |
| WO2023024407A1 (en) | Model pruning method and apparatus based on adjacent convolutions, and storage medium | |
| CN109271876B (en) | Video motion detection method based on time evolution modeling and multi-example learning | |
| CN110543906B (en) | Automatic skin recognition method based on Mask R-CNN model | |
| CN110321805B (en) | Dynamic expression recognition method based on time sequence relation reasoning | |
| CN114880925A (en) | Equipment life prediction method based on time convolution network and multi-layer self-attention | |
| CN103544499A (en) | Method for reducing dimensions of texture features for surface defect detection on basis of machine vision | |
| CN111898461A (en) | Time sequence behavior segment generation method | |
| Ni et al. | Light YOLO for high-speed gesture recognition | |
| CN104268507A (en) | Manual alphabet identification method based on RGB-D image | |
| CN112270405A (en) | Filter pruning method and system of convolution neural network model based on norm | |
| Yang et al. | Classification of medical images with synergic graph convolutional networks | |
| CN112215268A (en) | Method and device for classifying disaster weather satellite cloud pictures | |
| CN114781779A (en) | Unsupervised energy consumption abnormity detection method and device and storage medium | |
| CN111768803B (en) | General audio steganalysis method based on convolutional neural network and multitask learning | |
| CN107045624A (en) | A kind of EEG signals pretreatment rolled into a ball based on maximum weighted and sorting technique | |
| CN112380932B (en) | Vibration signal characteristic value selection method and elevator health state evaluation or fault diagnosis method | |
| CN112132207A (en) | Target detection neural network construction method based on multi-branch feature mapping | |
| CN112465054B (en) | FCN-based multivariate time series data classification method | |
| CN112541469B (en) | Crowd counting method and system based on self-adaptive classification | |
| CN110265151B (en) | Learning method based on heterogeneous temporal data in EHR |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20181123 |